It is expected from the market prediction data of eMarketer that the size of on-line retails market may go beyond 4.8 trillion by 2021, accounting for 17.5% of the total size of the global retail market. As the core carrier for global e-commerce, e-commerce platform has extended its business presence in more than 200 countries and regions in the world, playing the role of one of the most important participators and constructors for local consumer market. Moreover, various vertical and integrated e-commerce platforms are constantly emerging. 

It is well-known that performance marketing is one of the most significant channels for e-commerce enterprises in the acquisition of high-quality customers. E-commerce advertisers will spend large budgets in marketing and promotion campaigns in each month. In the actual operation process, however, advertisers usually come across circumstances like fund waste, budget overturns, imbalanced return in investment, etc. As a result, the issue of how to improve the return on investment (hereinafter short for “ROI”) of performance marketing is one of the principal pain points in front of most e-commerce advertisers.

After massive model training and performance tests, Yeahmobi has officially presented its in-depth solution of Predicted Payer. With the application of in-depth learning, the solution predicts the conversion possibility of visitors according to the user data and visit behaviors, and then applies the prediction results into the practical advertising process to improve the efficiency of advertising.

System Design

With the system architecture developed on Google Cloud Platform (hereinafter short for “GCP”), Predicted Payer adopts four modules of GCP in the design process, including BigQuery, Compute Engine, Storage and AutoML Tables.

BigQuery

Applicable for the storage, query and analysis of behavioral data

Compute Engine

Applied to realize the extraction of features and construction of training samples

Storage

Used for the storage of data in the intermediary processes

AutoML Tables

Serves for the training of automatic in-depth learning models

First of all, AutoML will train the in-depth learning model based on the personal information and visit behaviors of users on websites of clients collected by Google Analytics360 (“GA360” for short);

Second, the system will run the “Prediction Module” for new behavioral data of users in websites on each day to conduct in-depth evaluation for each user, and predict the possibility of effective conversion within the future 14 days;

Finally, identify users ready for payment through Google Ads, and arrange re-marketing campaign for such users to promote the rapid conversion.

Data Acquisition

Predicted Payer system collects and analyzes behavioral data of users on websites of clients through GA360, and arranges statistical analysis for the track of behaviors to acquire the status parameters and behavioral parameters of users.

Among them, status parameters include basic information, browser version, mobile platform, international geographic location and other relevant information of users. Behavioral parameters include statistics of the behaviors on the day, list of pages visited, behavior of adding any product into the cart, etc.

Among the 338 dimensions recorded for users, we extract 267 features that could decide the purchase behaviors of users for further model training.

 Model Training

As an automatic machine learning software developed by Google, AutoML Table could construct and deploy an advanced machine learning model for structural data automatically. The underlying software is able to integrate the framework and modules like Tensorflow to significantly shorten the lead time and reduce the difficulty of engineering development. As a result, the data scientists may concentrate more efforts on the specific business scenarios.

As shown in the following picture, we conducted the simulation training for data of client website in two weeks. The samples in the simulation training were above 2 million.

Performance Test

The above picture elaborates the working principle of combination between the prediction system of Predicted Payer developed based on machine learning and the advertising platform of Google Ads. The left side of the prediction system of Predicted Payer developed based on TensorFlow machine learning, and the right side is the system of Google Ads.

With processes like data acquisition, data clearing, data analysis and data prediction, the system screens out the most valuable users for clients, enlarging the quantity of seedling groups conveyed to Google Ads and shortening feedback time for the conversion from users to performance. As a result, the performance and results of machine learning and advertising of Google Ads are significantly improved. 

Currently, the in-depth solution of Predicted Payer has contributed to the success of many e-commerce advertisers in their advertising campaigns.

Segmented audience test was arranged on the basis of predicted grades of users. In the absence of any significant difference in CTR, we got the prediction result that the performance of the group with highest score improved for as high as five times in comparison with the performance of the group with lowest score. The application of predicted results from the overlapping of segment audiences in the optimization of ads could contribute to a CPA improvement from 40% to 60% for customers.

If you have any further concerns about the in-depth solution of Predicted Payer:

Please contact your Yeahmobi Account Manager 

Tel.：+86 18066732665  

Email: evan.yang@yeahmobi.com